Performance evaluation of algorithms for late potentials detection: A simulated study

Summary form only given. The identification of patients subject to life-threatening arrhythmias after myocardial infarction using signal-processing techniques for late potential (LP) detection in high-resolution surface ECGs has been investigated. Most of these methods are based on two steps: averaged ECG cycle construction and LP detection. The authors use simulation to study the intrinsic performance of four LP detection algorithms. An artificial noise-free ECG mean-cycle is generated (QRS power peak at 15 Hz) with floating-length QRS duration (91 q4 msec). Added noise is a Gaussian, stationary bandlimited, zero-mean process. Noise level was fixed at 1 mV rms. LPs were modeled as linear combinations of spikelike waveforms with changing morphology (band 5-250 Hz). LPs used were 50, 100, 150, and 200 ms long, starting from the R-wave peak. Time-domain detectors (TDD) show a high specificity, whereas frequency-domain detectors (FDDs) show a lower sensibility than TDDs. As expected, FDD sensibility decreases with LP decreasing, whereas TDDs based on signal rms tend to maintain a constant performance, except in very-short-duration LPs. TDDs based on QRS duration lack sensitivity for LP duration <150 msec.